Ranking data provide context but can also confuse

This dataviz from the Economist had me spending a lot of time clicking around - which means it is a success.

The graphic presents four measures of wellbeing in society - life expectancy, infant mortality rate, murder rate and prison population. The primary goal is to compare nations across those metrics. The focus is on comparing how certain nations (or subgroups) rank against each other, as indicated by the relative vertical position.

The Economist staff has a particular story to tell about racial division in the US. The dotted bars represent the U.S. average. The colored bars are the averages for Hispanic, white and black Americans. The wider the gap between the colored bars, the more variant is the experiences between American races.

The chart shows that the racial gap of life expectancy is the widest. For prison population, the U.S. and its racial subgroups occupy many of the lowest (i.e. least desirable) ranks, with the smallest gap in ranking.

***

The primary element of interactivity is hovering on a bar, which then highlights the four bars corresponding to the particular nation selected. Here is the picture for Thailand:

According to this view of the world, Thailand is a close cousin of the U.S. On each metric, the Thai value clings pretty near the U.S. average and sits within the range by racial groups. I'm surprised to learn that the prison population in Thailand is among the highest in the world.

Unfortunately, this chart form doesn't facilitate comparing Thailand to a country other than the U.S as one can highlight only one country at a time.

***

While the main focus of the chart is on relative comparison through ranking, the reader can extract absolute difference by reading the lengths of the bars.

This is a close-up of the bottom of the prison population metric:

The length of each bar displays the numeric data. The red line is an outlier in this dataset. Black Americans suffer an incarceration rate that is almost three times the national average. Even white Americans (blue line) is imprisoned at a rate higher than most countries around the world.

As noted above, the prison population metric exhibits the smallest gap between racial subgroups. This chart is a great example of why ranking data frequently hide important information. The small gap in ranking masks the extraordinary absolute difference in incareration rates between white and black America.

The difference between rank #1 and rank #2 is enormous.

The opposite situation appears for life expectancy. The life expectancy values are bunched up especially at the top of the scale. The absolute difference between Hispanic and black America is 82 - 75 = 7 years, which looks small because the axis starts at zero. On a ranking scale, Hispanic is roughly in the top 15% while black America is just above the median. The relative difference is huge.

For life expectancy, ranking conveys the view that even a 7-year difference is a big deal because the countries are tightly bunched together. For prison population, ranking shows the view that a multiple fold difference is "unimportant" because a 20-0 blowout and a 10-0 blowout are both heavy defeats.

***

Whenever you transform numeric data to ranks, remember that you are artificially treating the gap between each value and the next value as a constant, even when the underlying numeric gaps show wide variance.

 

 

 

 

 

Further exploration of tessellation density

Last year, I explored using bar-density (and pie-density) charts to illustrate 80/20-type distributions, which are very common in real life (link).

The key advantage of this design is that the most important units (i.e. the biggest stars/creators) are represented by larger pieces while the long tail is shown by little pieces. The skewness is encoded in the density of the tessellation.

So when the following chart showed up on my Twitter feed, I returned to the idea of using tessellation density as a visual cue.

This wbur chart is a good statistical chart - effiicient at communicating the data, but "boring". The only things I'd change is to remove the vertical axis, gridlines, and the decimals.

In concept, the underlying data is similar to the Youtube data. Less than 0.5 percent of Youtubers produced 38% of the views on the platform. The richest 1% of the population took 15% of Harvard's spots; the richest 20% took 70%.

As I explore this further, the analogy falls apart. In the Youtube scenario, the stars should naturally occupy bigger spaces. In the Harvard scenario, letting the children of the top 1% taking up more space on the chart doesn't really make sense since each incoming Harvard student has equal status.

Instead of going down that potential deadend, I investigated how tessellation density can be used for visualization. For one thing, tessellations are pretty things and appealing.

Here is something I created:

The chart is read vertically by comparing Harvard's selection of students with the hypothetical "ideal" of equal selection. (I don't agree that this type of equality is the right thing but let me focus on the visualization here.) This, selectivity is coded in the density. Selectivity is defined here as the over/under representation. Harvard is more "selective" in lower-income groups.

In the first and second columns, we see that Harvard's densities are lower than the densities as expected in the general population, indicating that the poorest 20%, and the middle 20% of the population are under-represented in Harvard's student body. Then in the third column, the comparison flips. The density in the top box is about 3-4 times as high as the bottom box. You may have to expand the graphic to see the 1% slither, which also shows a much higher density in the top box.

I was surprised by how well I was able to eyeball the relative densities. You can try it and let me know how you fare.

(There is even a trick to do this. From the diagram with larger pieces, pick a representative piece. Then, roughly estimate how many smaller pieces from the other tessellation can fit into that representative piece. Using this guideline, I estimate that the ratios of the densities to be 1:6, 1:2, 3:1, 10:1. The actual ratios are 1:6.7, 1:2.5, 3:1, 15:1. I find that my intuition gets me most of the way there even if I don't use this trick.)

Density encoding is under-used as a visual cue. I think our ability to compare densities is surprisingly good (when the units are not overlapping). Of course, you wouldn't use density if you need to be precise, just as you wouldn't use color, or circular areas. Nevertheless, there are many occasions where you can afford to be less precise, and you'd like to spice up your charts.

Plotting the signal or the noise

Antonio alerted me to the following graphic that appeared in the Economist. This is a playful (?) attempt to draw attention to racism in the game of football (soccer).

The analyst proposed that non-white players have played better in stadiums without fans due to Covid19 in 2020 because they have not been distracted by racist abuse from fans, using Italy's Serie A as the case study.

The chart struggles to bring out this finding. There are many lines that criss-cross. The conclusion is primarily based on the two thick lines - which show the average performance with and without fans of white and non-white players. The blue line (non-white) inched to the right (better performance) while the red line (white) shifted slightly to the left.

If the reader wants to understand the chart fully, there's a lot to take in. All (presumably) players are ranked by the performance score from lowest to highest into ten equally sized tiers (known as "deciles"). They are sorted by the 2019 performance when fans were in the stadiums. Each tier is represented by the average performance score of its members. These are the values shown on the top axis labeled "with fans".

Then, with the tiers fixed, the players are rated in 2020 when stadiums were empty. For each tier, an average 2020 performance score is computed, and compared to the 2019 performance score.

The following chart reveals the structure of the data:

The players are lined up from left to right, from the worst performers to the best. Each decile is one tenth of the players, and is represented by the average score within the tier. The vertical axis is the actual score while the horizontal axis is a relative ranking - so we expect a positive correlation.

The blue line shows the 2019 (with fans) data, which are used to determine tier membership. The gray dotted line is the 2020 (no fans) data - because they don't decide the ranking, it's possible that the average score of a lower tier (e.g. tier 3 for non-whites) is higher than the average score of a higher tier (e.g. tier 4 for non-whites).

What do we learn from the graphic?

It's very hard to know if the blue and gray lines are different by chance or by whether fans were in the stadium. The maximum gap between the lines is not quite 0.2 on the raw score scale, which is roughly a one-decile shift. It'd be interesting to know the variability of the score of a given player across say 5 seasons prior to 2019. I suspect it could be more than 0.2. In any case, the tiny shifts in the averages (around 0.05) can't be distinguished from noise.

***

This type of analysis is tough to do. Like other observational studies, there are multiple problems of biases and confounding. Fan attendance was not the only thing that changed between 2019 and 2020. The score used to rank players is a "Fantacalcio algorithmic match-level fantasy-football score." It's odd that real-life players should be judged by their fantasy scores rather than their on-the-field performance.

The causal model appears to assume that every non-white player gets racially abused. At least, the analyst didn't look at the curves above and conclude, post-hoc, that players in the third decile are most affected by racial abuse - which is exactly what has happened with the observational studies I have featured on the book blog recently.

Being a Serie A fan, I happen to know non-white players are a small minority so the error bars are wider, which is another issue to think about. I wonder if this factor by itself explains the shifts in those curves. The curve for white players has a much higher sample size thus season-to-season fluctuations are much smaller (regardless of fans or no fans).

 

 

 

 

Illustrating differential growth rates

Reader Mirko was concerned about a video published in Germany that shows why the new coronavirus variant is dangerous. He helpfully provided a summary of the transcript:

The South African and the British mutations of the SARS-COV-2 virus are spreading faster than the original virus. On average, one infected person infects more people than before. Researchers believe the new variant is 50 to 70 % more transmissible.

Here are two key moments in the video:

This seems to be saying the original virus (left side) replicates 3 times inside the infected person while the new variant (right side) replicates 19 times. So we have a roughly 6-fold jump in viral replication.

Later in the video, it appears that every replicate of the old virus finds a new victim while the 19 replicates of the new variant land on 13 new people, meaning 6 replicates didn't find a host.

As Mirko pointed out, the visual appears to have run away from the data. (In our Trifecta Checkup, we have a problem with the arrow between the D and the V corners. What the visual is saying is not aligned with what the data are saying.)

***

It turns out that the scientists have been very confusing when talking about the infectiousness of this new variant. The most quoted line is that the British variant is "50 to 70 percent more transmissible". At first, I thought this is a comment on the famous "R number". Since the R number around December was roughly 1 in the U.K, the new variant might bring the R number up to 1.7.

However, that is not the case. From this article, it appears that being 5o to 70 percent more transmissible means R goes up from 1 to 1.4. R is interpreted as the average number of people infected by one infected person.

Mirko wonders if there is a better way to illustrate this. I'm sure there are many better ways. Here's one I whipped up:

The left side is for the 40% higher R number. Both sides start at the center with 10 infected people. At each time step, if R=1 (right side), each of the 10 people infects 10 others, so the total infections increase by 10 per time step. It's immediately obvious that a 40% higher R is very serious indeed. Starting with 10 infected people, in 10 steps, the total number of infections is almost 1,000, almost 10 times higher than when R is 1.

The lines of the graphs simulate the transmission chains. These are "average" transmission chains since R is an average number.

 

P.S. [1/29/2021: Added the missing link to the article in which it is reported that 50-70 percent more transmissible implies R increasing by 40%.]

 

 

Atypical time order and bubble labeling

This chart appeared in a Charles Schwab magazine in Summer, 2019.

This bubble chart does not print any data labels. The bubbles take our attention but the designer realizes that the actual values of the volatility are not intuitive numbers. The same is true of any standard deviation numbers. If you're told SD of a data series is 3, it doesn't tell you much by itself.

I first transformed this chart into the equivalent column chart:

Two problems surface on the axes.

For the time axis, the years are jumbled. Readers experience vertigo, as we try to figure out how to read the chart. Our expectation that time moves left to right is thwarted. This ordering also requires every single year label to be present.

For the vertical axis, I could have left out the numbers completely. They are not really meaningful. These represent the areas of the bubbles but only relative to how I measured them.

***

In the next version, I sorted time in the conventional manner. Following Tufte's classic advice, only the tops of the columns are plotted.

What you see is that this ordering is much easier to comprehend. Figuring out that 2018 is an average year in terms of volatility is not any harder than in the original. In fact, we can reproduce the order of the previous chart just by letting our eyes sweep top to bottom.

To make it even easier to read the vertical axis, I converted the numbers into an index, with the average volatility as 100 %  (assigned to 0% on the chart) .

Now, you can see that 2018 is roughly at the average while 2008 is 400% above the average level. (How should we interpret this statement? That's a question I pose to my statistics students. It's not intuitive how one should interpret the statement that the standard deviation is 5 times higher.)

 

 

Convincing charts showing containment measures work

The disorganized nature of U.S.'s response to the coronavirus pandemic has created a sort of natural experiment that allows data journalists to explore important scientific questions, such as the impact of containment measures on cases and hospitalizations. This New York Times article represents the best of such work.

The key finding of the analysis is beautifully captured by this set of scatter plots:

Each dot is a state. The cases (left plot) and hospitalizations (right plot) are plotted against the severity of containment measures for November. The negative correlation is unmistakable: the more containment measures taken, the lower the counts.

There are a few features worth noting.

The severity index came from a group at Oxford, and is a number between 0 and 100. The journalists decided to leave out the numerical labels, instead simply showing More and Fewer. This significantly reduces processing time. Readers won't be able to understand the index values anyway without reading the manual.

The index values are doubly encoded. They are first encoded by the location on the horizontal axis and redundantly encoded on the blue-red scale. Ordinarily, I do not like redundant encoding because the reader might assume a third dimension exists. In this case, I had no trouble with it.

The easiest way to see the effect is to ignore the muddy middle and focus on the two ends of the severity index. Those states with the fewest measures - South Dakota, North Dakota, Iowa - are the worst in cases and hospitalizations while those states with the most measures - New York, Hawaii - are among the best. This comparison is similar to what is frequently done in scientific studies, e.g. when they say coffee is good for you, they typically compare heavy drinkers (4 or more cups a day) with non-drinkers, ignoring the moderate and light drinkers.

Notably, there is quite a bit of variability for any level of containment measures - roughly 50 cases per 100,000, and 25 hospitalizations per 100,000. This indicates that containment measures are not sufficient to explain the counts. For example, the hospitalization statistic is affected by the stock of hospital beds, which I assume differ by state.

Whenever we use a scatter plot, we run the risk of xyopia. This chart form invites readers to explain an outcome (y-axis values) using one explanatory variable (on x-axis). There is an assumption that all other variables are unimportant, which is usually false.

***

Because of the variability, the horizontal scale has meaningless precision. The next chart cures this by grouping the states into three categories: low, medium and high level of measures.

This set of charts extends the time window back to March 1. For the designer, this creates a tricky problem - because states adapt their policies over time. As indicated in the subtitle, the grouping is based on the average severity index since March, rather than just November, as in the scatter plots above.

***

The interplay between policy and health indicators is captured by connected scatter plots, of which the Times article included a few examples. Here is what happened in New York:

Up until April, the policies were catching up with the cases. The policies tightened even after the case-per-capita started falling. Then, policies eased a little, and cases started to spike again.

The Note tells us that the containment severity index is time shifted to reflect a two-week lag in effect. So, the case count on May 1 is not paired with the containment severity index of May 1 but of April 15.

***

You can find the full article here.

 

 

 

Bloomberg made me digest these graphics slowly

Ask the experts to name the success metric of good data visualization, and you will receive a dozen answers. The field doesn't have an all-encompassing metric. A useful reference is Andrew Gelman and Antony Urwin (2012) in which they discussed the tradeoff between beautiful and informative, which derives from the familiar tension between art and science.

For a while now, I've been intrigued by metrics that measure "effort". Some years ago, I described the concept of a "return on effort" in this post. Such a metric can be constructed like the dominating financial metric of return on investment. The investment here is an investment of time, of attention. I strongly believe that if the consumer judges a data visualization to be compelling, engaging or  ell constructed, s/he will expend energy to devour it.

Imagine grub you discard after the first bite, compared to the delicious food experienced slowly, savoring every last bit.

I'm writing this post while enjoying the September issue of Bloomberg Businessweek, which focuses on the upcoming U.S. Presidential election. There are various graphics infused into the pages of the magazine. Many of these graphics operate at a level of complexity above what typically show up in magazines, and yet I spent energy learning to understand them. This response, I believe, is what visual designers should aim for.

***

Today, I discuss one example of these graphics, shown on the right. You might be shocked by the throwback style of these graphics. They look like they arrived from decades ago!

Grayscale, simple forms, typewriter font, all caps. Have I gone crazy?

The article argues that a town like Ambridge in Beaver County, Pennslyvania may be pivotal in the November election. The set of graphics provides relevant data to understand this argument.

It's evidence that data visualization does not need whiz-bang modern wizardry to excel.

Let me focus on the boxy charts from the top of the column. These:

These charts solve a headache with voting margin data in the U.S.  We have two dominant political parties so in any given election, the vote share data split into three buckets: Democratic, Republican, and a catch-all category that includes third parties, write-ins, and none of the above. The third category rarely exceeds 5 percent.  A generic pie chart representation looks like this:

Stacked bars have this look:

In using my Trifecta framework (link), the top point is articulating the question. The primary issue here is the voting margin between the winner and the second-runner-up, which is the loser in what is typically a two-horse race. There exist two sub-questions: the vote-share difference between the top two finishers, and the share of vote effectively removed from the pot by the remaining candidates.

Now, take another look at the unusual chart form used by Bloomberg:

The catch-all vote share sits at the bottom while the two major parties split up the top section. This design demonstrates a keen understanding of the context. Consider the typical outcome, in which the top two finishers are from the two major parties. When answering the first sub-question, we can choose the raw vote shares, or the normalized vote shares. Normalizing shifts the base from all candidates to the top two candidates.

The Bloomberg chart addresses both scales. The normalized vote shares can be read directly by focusing only on the top section. In an even two-horse race, the top section is split by half - this holds true regardless of the size of the bottom section.

This is a simple chart that packs a punch.

 

Election visuals 2: informative and playful

In yesterday's post, I reviewed one section of 538's visualization of its election forecasting model, specifically, the post focuses on the probability plot visualization.

The visualization, technically called  a pdf, is a mainstay of statistical graphics. While every one of 40,000 scenarios shows up on this chart, it doesn't offer a direct answer to our topline question. What is Nate's call at this point in time? Elsewhere in their post, we learn that the 538 model currently gives Biden a 75% chance of winning, thrice that of Trump's.

In graphical terms, the area to the right of the 270-line is three times the size of the left area (on the bottom chart). That's not apparent in the pdf representation. Addressing this, statisticians may convert the pdf into a cdf, which depicts the cumulative area as we sweep from the left to the right along the horizontal axis.  

The cdf visualization rarely leaves the pages of a scientific journal because it's not easy for a novice to understand. Not least because the relevant probability is 1 minus the cumulative probability. The cdf for the bottom chart will show 25% at the 270-line while the chance of Biden winning is 1 - 25% = 75%.

The cdf presentation is also wasteful for the election scenario. No one cares about any threshold other than the 270 votes needed to win, but the standard cdf shows every possible threshold.

The second graphical concept in the 538 post (link) is an attempt to solve this problem.

If you drop all the dots to an imaginary horizontal baseline, the above dotplot looks like this:

There is a recent trend toward centering dots to produce symmetry. It's actually harder to perceive the differences in heights of the band.

The secret sauce is to put down 100 dots, with a 75-25 blue-red split that conveys the 75% chance of a Biden win. Imposing the pdf line from the other visualization, I find that the density of dots roughly mimics the probability of outcomes.

It's easier to estimate the blue vs red areas using those dots than the lines.

The dots are stuffed toys. Clicking on each dot reveals a map showing one of the 40,000 scenarios. It displays which candidate wins which state. For example, the most extreme example of a Trump win is:

Here is a scenario of a razor-tight election won by Trump:

This presentation has a weakness as well. It gives the impression that each of the dots is equally important because they are the same size. In reality, the importance of each dot is proportional to the height of the band. Since the band is generally wider near the middle, the dots near the middle are more likely scenarios than the dots shown on the two edges.

On balance, I like this visualization that is both informative and playful.

As before, what strikes me about the simulation result is the flatness of the probability surface. This feature is obscured when we summarize the result as 75% chance of a Biden victory.

Election visuals: three views of FiveThirtyEight's probabilistic forecasts

As anyone who is familiar with Nate Silver's forecasting of U.S. presidential elections knows, he runs a simulation that explores the space of possible scenarios. The polls that provide a baseline forecast make certain assumptions, such as who's a likely voter. Nate's model unshackles these assumptions from the polling data, exploring how the outcomes vary as these assumptions shift.

In the most recent simulation, his computer explores 40,000 scenarios, each of which predicts a split of the electoral vote, from which the winner of the election can be determined. The model's outcome is usually summarized by a winning probability, which is just the proportion of scenarios under which one candidate wins.

This type of forecasting was responsible for the infamous meltdown in 2016 when most of these models - Nate's being an exception - issued extremely confident predictions that Hillary Clinton wins with 95% or higher probability. Essentially, the probability distribution collapses to a point. This is analogous to an extremely narrow confidence band, indicating almost zero uncertainty about the event. It was as if almost all of the 40,000 scenarios predicted Clinton to be the winner.

The 538 data team has come up with various ways of visualizing the outputs of the model (link). The entire post is worth reading. Here, I'll highlight the most scientific, and direct visual representation, which is the third display.

We start by looking at the bottom of the two charts, showing the predicted electoral votes won  by Democratic challenger Joe Biden, in each of the 40,000 scenarios. Our attention is directed to the thick line that gives the relative chance of Biden's electoral-vote tally. This line is a smoothed summary of the columns in the background, which show the number of times the simulation produces each electoral-vote count.

The highlighted, right side of the chart recounts scenarios in which Biden becomes President, that is to say, he wins more than 270 electoral votes (out of 538, doh). The faded, left side represents scenarios in which Biden is defeated and Trump wins a second term.

The reason I focused on the bottom chart is that the top chart is merely a mirror image of this one. Just reflect the bottom chart around the vertical axis of 270 electoral votes, change the color scheme to red, and swap annotations related to Trump and Biden, and you get the other chart. This is because the narrative has excluded third-party and write-in candidates, leaving us with a zero-sum situation.

Alternatively, one can jam both charts into one, while supplying extra labels, like this:

I prefer the denser single chart because my mind wanders away searching for extra meaning when chart elements are mirrored.

One advantage of the mirrored presentation is that the probability profiles of the potential Trump or Biden wins can be directly compared. We learn that Trump's winning margins are smaller, rarely above 150, and never above 250.

This comparison is made easier by flipping left side of the chart onto the right side:

Those are three different visualizations using the same chart form. I'd have to run a poll to figure out which is the best. What's your opinion?

Deaths as percent neither of cases nor of population. Deaths as percent of normal.

Yesterday, I posted a note about excess deaths on the book blog (link). The post was inspired by a nice data visualization by the New York Times (link). This is a great example of data journalism.

Excess deaths is a superior metric for measuring the effect of Covid-19 on public health. It's better than deaths as percent of cases. Also better than percent of the population.What excess deaths measure is deaths as a percent of normal. Normal is usually defined as the average deaths in the respective week in years past.

The red areas indicate how far the deaths in the Southern states are above normal. The highest peak, registered in Texas in late July, is 60 percent above the normal level.

***

The best way to appreciate the effort that went into this graphic is to imagine receiving the outputs from the model that computes excess deaths. A three-column spreadsheet with columns "state", "week number" and "estimated excess deaths".

The first issue is unequal population sizes. More populous states of course have higher death tolls. Transforming death tolls to an index pegged to the normal level solves this problem. To produce this index, we divide actual deaths by the normal level of deaths. So the spreadsheet must be augmented by two additional columns, showing the historical average deaths and actual deaths for each state for each week. Then, the excess death index can be computed.

The journalist builds a story around the migration of the coronavirus between different regions as it rages across different states  during different weeks. To this end, the designer first divides the dataset into four regions (South, West, Midwest and Northeast). Within each region, the states must be ordered. For each state, the week of peak excess deaths is identified, and the peak index is used to sort the states.

The graphic utilizes a small-multiples framework. Time occupies the horizontal axis, by convention. The vertical axis is compressed so that the states are not too distant. For the same reason, the component graphs are allowed to overlap vertically. The benefit of the tight arrangement is clearer for the Northeast as those peaks are particularly tall. The space-saving appearance reminds me of sparklines, championed by Ed Tufte.

There is one small tricky problem. In most of June, Texas suffered at least 50 percent more deaths than normal. The severity of this excess death toll is shortchanged by the low vertical height of each component graph. What forced such congestion is probably the data from the Northeast. For example, New York City:

 

New York City's death toll was almost 8 times the normal level at the start of the epidemic in the U.S. If the same vertical scale is maintained across the four regions, then the Northeastern states dwarf all else.

***

One key takeaway from the graphic for the Southern states is the persistence of the red areas. In each state, for almost every week of the entire pandemic period, actual deaths have exceeded the normal level. This is strong indication that the coronavirus is not under control.

In fact, I'd like to see a second set of plots showing the cumulative excess deaths since March. The weekly graphic is better for identifying the ebb and flow while the cumulative graphic takes measure of the total impact of Covid-19.

***

The above description leaves out a huge chunk of work related to computing excess deaths. I assumed the designer receives these estimates from a data scientist. See the related post in which I explain how excess deaths are estimated from statistical models.